diff-artifact

name	diff-artifact
description	Publishes an annotated diff walkthrough as a claude.ai artifact. Scrubs for secrets, then builds a self-contained page with per-hunk reviewer notes. Use when asked to publish or share a diff.
allowed-tools	Bash, Read, Write, Skill, Artifact, AskUserQuestion, ToolSearch, ExitPlanMode

diff-artifact

Turn a branch diff, commit range, or pull request into a single annotated walkthrough page published to claude.ai — the reviewer sees not just what changed, but why each hunk changed and what to look at first.

Overview

A raw git diff tells a reviewer what moved; it never tells them what matters. This skill reads the diff, writes a reviewer note against each meaningful hunk, and publishes the result as one self-contained artifact page with a sticky file sidebar, severity labels, and adaptive light/dark theming.

Publishing sends code to an external service, so the security-scrub gate runs before any publish and a BLOCKED verdict is a hard stop.

Exit plan mode

ExitPlanMode is a deferred tool. Only call it if currently in plan mode — skip this step entirely when not in plan mode. When calling: use ToolSearch with select:ExitPlanMode first, then call ExitPlanMode silently.

Step 1 — Resolve the diff source

Pick the first mode that matches the user's argument:

Mode	Trigger	Command
Branch (default)	no argument	git diff <default-branch>...HEAD
Range	abc123..def456	git diff <range>
PR	#42 or a PR URL	gh pr diff 42

Resolve the default branch rather than assuming main:

DEFAULT_BRANCH=$(git symbolic-ref --quiet refs/remotes/origin/HEAD 2>/dev/null | sed 's|.*/||')
DEFAULT_BRANCH=${DEFAULT_BRANCH:-main}
git diff "${DEFAULT_BRANCH}...HEAD"

PR-mode degradation. PR mode needs both the gh CLI and a GitHub remote. When either is missing, do not fail and do not merely report it — say so plainly and actually produce the branch diff. Every degradation path must still end with a diff on disk:

DIFF_FILE="<scratchpad>/diff.patch"

use_pr=0
if [ -n "${PR:-}" ]; then
  if gh auth status >/dev/null 2>&1 &&
     git remote get-url origin 2>/dev/null | grep -qi 'github\.com'; then
    use_pr=1
  else
    echo "PR mode unavailable (gh missing/unauthenticated, or origin is not GitHub) — using branch mode"
  fi
fi

if [ "$use_pr" = 1 ]; then
  gh pr diff "$PR" > "$DIFF_FILE" || { echo "gh pr diff failed — using branch mode"; use_pr=0; }
fi
[ "$use_pr" = 1 ] || git diff "${DEFAULT_BRANCH}...HEAD" > "$DIFF_FILE"

If $DIFF_FILE is empty, tell the user there is nothing to publish and stop.

Step 2 — Scrub before anything else (blocking gate)

Write the diff to a scratch file and run the scrub over it via the Skill tool (social-media-tools:security-scrub). This gate is blocking, not advisory — publishing is external sharing.

• GATE RESULT: BLOCKED → hard stop. Do not publish, do not write the page, do not offer an override. Report the masked findings and stop.
• GATE RESULT: CANCELLED → the user declined. Stop.
• GATE RESULT: APPROVED → continue to Step 3.

If security-scrub is unavailable (social-media-tools not installed), do not silently skip the gate — tell the user the scan could not run and ask via AskUserQuestion whether to continue with an unscanned diff.

Step 3 — Annotate the hunks

Read the changed files for context where the diff alone is ambiguous, then write one note per meaningful hunk. Each note gets a severity:

Scrub coverage. Step 2 scanned the diff. Reading surrounding file context here can pull in text the diff never contained, so anything quoted from a file is outside that scan. Quote sparingly, and note that Step 5 rescans the rendered page — that second gate, not this one, is what covers annotations.

Severity	Meaning
critical	Reviewer must look here — correctness, security, data loss
warn	Worth a second opinion — edge cases, unclear intent
note	Context only — renames, formatting, mechanical churn

Explain the reasoning, not the syntax. "Guards against the empty-array case that crashed the importer" is a note; "adds an if statement" is noise.

Cap-and-summarize. Artifacts are capped at 16 MiB rendered. Annotate at most 20 files and 8 hunks per file. Every file beyond that budget renders as a one-line summary row in the sidebar and body (path — +12/−3, not annotated) rather than a full diff. Tell the user in the final report how many files were summarized rather than annotated, so a truncated review is never mistaken for a complete one.

Step 4 — Build the page

Load the artifact-design skill first to calibrate design investment, then Write one self-contained .html file to the scratchpad. Requirements:

• Self-contained — a strict CSP blocks every external request. Inline all CSS; no CDN links, no web fonts, no remote images, no fetch.
• Single page — in-page anchors only (#file-3); relative links break.
• Sticky file sidebar — every changed file with add/del counts, anchored to its section. Summarized-only files appear here too, visibly marked.
• Severity legend — pair every color with a text label. Color alone fails colorblind readers; the label is what carries the meaning.
• Adaptive theme — @media (prefers-color-scheme: dark) for both palettes.
• Escape diff content — &, <, > in code become entities. An unescaped diff of HTML silently destroys the page.
• Title — read ${CLAUDE_PLUGIN_ROOT}/references/titles.md and set the <title> by its rules. The diff is in context here; the subject is the theme the changed files share, never the user's phrasing of the request.
• Write the page content only (no <!doctype>/<html>/<head>/<body> — those are added at publish time).

Step 5 — Gate the rendered page (size, then scrub)

Both checks run on the rendered HTML, not on the inputs.

Size. The file/hunk budget bounds how many hunks render, but a single generated file can carry one enormous hunk, so the budget alone does not enforce the cap. Measure the page and shrink it until it fits:

CAP=$((16 * 1024 * 1024))   # 16 MiB rendered-artifact limit
while [ "$(wc -c < "$SCRATCH_HTML")" -ge "$CAP" ]; do
  # Demote the largest annotated file to a one-line summary row, re-render, re-measure.
  echo "Page over 16 MiB — demoting the largest annotated file to a summary row"
done

Count every demotion here on top of the Step 3 budget, and report the total.

Scrub. Rescan the finished page with social-media-tools:security-scrub. Step 2 covered the diff; this covers everything the page actually publishes — annotations, quoted file context, titles. Same verdicts, same blocking behaviour: BLOCKED is a hard stop with no override, CANCELLED stops, only APPROVED proceeds.

Step 6 — Save the durable copy

Write the page to the .claude/artifacts/ inbox before publishing — the scratchpad is temporary, so a page that only ever lived there is not a fallback. The URL is recorded later, only if a publish actually succeeds.

Key the filename by what the diff is, never by date. A date in the key means tomorrow's run of the same branch misses today's artifact-url: and silently mints a second page:

mkdir -p .claude/artifacts
# Deterministic key: pr-42 | range-abc123-def456 | branch-<slug>
case "$MODE" in
  pr)     key="pr-${PR}" ;;
  range)  key="range-$(echo "$RANGE" | tr '.' '-')" ;;
  *)      key="branch-$(git rev-parse --abbrev-ref HEAD | tr '/' '-')" ;;
esac
target=".claude/artifacts/diff-${key}.html"
cp "$SCRATCH_HTML" "$target"

Step 7 — Publish, then record the URL

Artifact is a deferred tool: use ToolSearch with select:Artifact first.

Before publishing, read $target's first line: if it carries an <!-- artifact-url: ... --> comment from an earlier run, pass that URL to Artifact's url parameter so the same page updates. Without it every session mints a new URL and the link you already shared goes stale.

Publish $target with a one-sentence description and the favicon 🔍, kept identical across republishes — users find the tab by its icon.

On success, record the URL into the durable copy so later sessions find it:

tmp=$(mktemp)
{ echo "<!-- artifact-url: $URL -->"; grep -v '^<!-- artifact-url:' "$target"; } > "$tmp"
mv "$tmp" "$target"

Report the claude.ai URL, the local path, and how many files were summarized or demoted — a truncated review must never read as a complete one.

On publish failure (no claude.ai login, or publishing unavailable): this is not an edge case — sharing beyond the author needs Team/Enterprise, so on Pro and Max the fallback is how the page reaches teammates. $target already holds the page, with no artifact-url: line since nothing was published. Say plainly that publishing did not happen and why, give the path, and offer social-media-tools:save-artifact to publish it into the repo's GitHub Pages artifacts gallery instead. Never report a URL that a publish did not return.